This looks really similar (in concept, your design is way better and you have more features) to something I built with my lawyer/legaltech business partner a few years ago. We even applied to YC and didn’t get in.
I’m curious, though, how accurate is your llm for things like trying to “pull subtle context from an attorney’s browser activity to associate that work with a client”? I suspect this needs to be damn near bullet proof, otherwise attorneys will tear it apart. We stuck purely to file tracking, comparing diffs and syncing activity in real time on a an app with GitHub like heatmap that could also flag suspicious time entries. It worked great but we had a hard time getting adoption because people did not want to force their attorneys to use it. Interesting to see the times change dynamics, good luck!
Really cool that you've seen this space up close! It's definitely a hard problem. With recent improvements to these large models, we're able to get solid accuracy (which, as you alluded to, is very important). Often, users will minimally revise the narrative descriptions before submitting.
I’m curious, though, how accurate is your llm for things like trying to “pull subtle context from an attorney’s browser activity to associate that work with a client”? I suspect this needs to be damn near bullet proof, otherwise attorneys will tear it apart. We stuck purely to file tracking, comparing diffs and syncing activity in real time on a an app with GitHub like heatmap that could also flag suspicious time entries. It worked great but we had a hard time getting adoption because people did not want to force their attorneys to use it. Interesting to see the times change dynamics, good luck!